Resilience Architecture Principles
- Non-Blocking Task Dispatch: Fast distributed Celery and Redis event loops prevent long-running inference jobs from starving short-lived webhook deliveries.
- Liveness Heartbeats: Workers report that they are alive on a schedule; a missed heartbeat raises an alert so the job can be recovered.
- Graceful Degradation Fallbacks: If the main LLM provider fails or slows down, requests fall back to a second provider.
